基于无监督学习的抽油机井示功图自动聚类与批量标注方法  

Automatic clustering and batch marking method for indicator diagram of pumping well based on unsupervised learning

作　　者：王相[1] 邵志伟 张雷 张中慧 肖姝 WANG Xiang;SHAO Zhiwei;ZHANG Lei;ZHANG Zhonghui;XIAO Shu(School of Petroleum and Natural Gas Engineering,Changzhou University,Changzhou,Jiangsu 213000,China;Sinopec Shengli Oilfield Branch Petroleum Engineering Technology Research Institute,Dongying,Shandong 257015,China)

机构地区：[1]常州大学石油与天然气工程学院,江苏常州213000 [2]中国石化胜利油田分公司石油工程技术研究院,山东东营257015

出　　处：《中国科技论文》2024年第1期63-69,共7页China Sciencepaper

基　　金：国家自然科学基金资助项目(52204027);江苏省研究生科研与实践创新计划项目(KYCX23_3146)。

摘　　要：为充分利用大量未标注样本、节约人力与时间,提出了基于无监督学习的抽油机井示功图自动聚类与批量标注方法。首先,将抽油机驴头往复运动产生的位移、载荷数据转化为示功图图片样本,其中,示功图的横坐标为位移,纵坐标为载荷;其次,加载在ImageNet上训练过的带有一系列权重参数、具有强特征提取能力的卷积神经网络模型;然后,去除该网络模型的全连接层,利用该网络模型提取示功图图片样本的特征;最后,利用k-means聚类算法对提取到的特征进行聚类分析,将具有相似特征的示功图聚到同一文件夹中。批量的对示功图聚类结果进行快速标注,从而形成抽油机井故障诊断的示功图样本集。实验随机搜集了100口抽油机井的20 000条示功图数据,结果表明,基于无监督学习的抽油机井示功图自动聚类与批量标注方法耗时短、准确率高,为示功图样本集标注提供了一种高效方法,对于充分挖掘油田大数据的应用价值具有示范意义。In order to make full use of a large number of unmarked samples and save manpower and time,an automatic clustering and batch marking method for indicator diagram of pumping well based on unsupervised learning was proposed.First,the displace-ment and load data generated by the reciprocating motion of the pumping unit horsehead were converted into the sample of the indica-tor diagram,where the abscissa of the indicator diagram was the displacement and the ordinate is the load.Secondly,the convolu-tion neural network model with a series of weight parameters and strong feature extraction ability that had been trained on ImageNet was loaded.Then,the full connection layer of the network model was removed,and the network model to extract the characteristics of indicator diagram image samples was used.Finally,k-means clustering algorithm was used to cluster the extracted features and cluster the indicator diagrams with similar features into the same folder.Batch of indicator diagram clustering results were quickly marked to form a sample set of indicator diagrams for fault diagnosis of pumping wells.Twenty thousand indicator diagram data from 100 pumping wells were randomly collected.The results show that the automatic clustering and batch marking method for indicator diagram of pumping well based on unsupervised learning is time-efficient and highly accurate.This method provides an efficient method for indicator sample set marking,which has exemplary significance for fully mining the application value of oilfield big data.

关 键 词：抽油机 示功图 故障诊断 K-MEANS聚类 样本标注 

分 类 号：TE355[石油与天然气工程—油气田开发工程]